Optimization problems play a pivotal role in the development of various fields. Among them, mixed-variable optimization problems (MVOPs) involving both continuous and discrete variables warrant particular attention, as they are prevalent in engineering design and pose substantial computational challenges for solution methods. To this end, this paper proposes a novel variant of differential evolution (DE), termed the model archive and ensemble learning-based differential evolution algorithm (MEDE). Specifically, a data enhancing strategy based on dynamic opposition-based learning is first devised to improve the diversity of the dominant population. Furthermore, a model archive is designed to store the probabilistic distribution models learned during the early-stage evolutionary process, which enables the algorithm to efficiently exploit learned probabilistic distribution models while preventing premature convergence. Finally, by utilizing stacking ensemble learning, the model’s predictive accuracy and generalization capability have been enhanced, thereby improving the overall optimization performance of the MEDE algorithm. To validate the effectiveness and efficiency of the proposed algorithm, 28 artificial benchmark functions are selected to conduct performance evaluation experiments. The experimental results clearly show that compared with other similar algorithms, i.e., AEDA, S-PSO, \(ACO_{mv}\) , and \(DE_{mv}\) , the proposed algorithm shows significant advantages in terms of quality and efficiency.